Intelligent characteristic evolution analysis method and system for time sequence path

Abstract

The invention provides an intelligent feature evolution analysis method and system for a time sequence path, and relates to the technical field of feature evolution, and the method comprises the steps: obtaining time sequence path data, and carrying out the preprocessing to generate a standardized feature sequence; performing Monte Carlo sampling and transition probability matrix calculation to obtain a characteristic evolution strength index; setting an evaluation time window by adopting a beam search algorithm, calculating a stability score, and selecting an optimal path branch; and finally, determining an optimal characteristic evolution path based on a weighted combination of the evolution strength index and the stability score, and generating an analysis report. According to the method, the evolution law of key features in the time series data can be effectively identified, and the analysis accuracy and efficiency are improved.

Description

Intelligent Feature Evolution Analysis Method and System for Temporal Paths Technical Field

[0001] This invention relates to feature evolution technology, and more particularly to a method and system for intelligent feature evolution analysis of temporal paths. Background Technology

[0002] Temporal path feature evolution analysis is an important research area in data science and artificial intelligence, with wide applications in fields such as financial market analysis, user behavior prediction, traffic flow monitoring, and disease transmission pattern research. Temporal path data typically contains feature information that changes over time. By analyzing the evolutionary patterns of these features, potential patterns and trends can be discovered, providing important basis for decision-making.

[0003] Traditional time-series path feature analysis methods primarily rely on fundamental statistical and machine learning techniques, such as time series analysis, hidden Markov models, and recurrent neural networks. These methods learn from historical data to build models that predict future feature evolution paths. With the advancement of big data technology and computing power, time-series path feature analysis has evolved from simple linear prediction to complex nonlinear dynamic system modeling.

[0004] Existing methods lack robustness when processing high-dimensional, high-noise time-series path data, struggling to effectively distinguish between real signals and random noise, resulting in low accuracy in identifying feature evolution paths. Traditional analysis methods lack effective quantification of uncertainties during feature evolution, failing to fully capture the probability distribution characteristics of feature transitions at different time points, thus affecting the reliability of feature evolution trend prediction. Current technologies lack adaptive evaluation mechanisms to dynamically adjust the search strategy for feature evolution paths, making it difficult to maintain computational efficiency while ensuring accuracy in feature evolution analysis, particularly performing poorly in long-term series feature evolution analysis. Summary of the Invention

[0005] The present invention provides a method and system for intelligent feature evolution analysis of time-series paths, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an intelligent feature evolution analysis method for time-series paths, comprising: acquiring time-series path data; preprocessing the time-series path data to generate a standardized feature sequence; generating multiple candidate feature values ​​at each time point using Monte Carlo sampling based on the standardized feature sequence; calculating a transition probability matrix between adjacent time points for the candidate feature values; iteratively optimizing the transition probability matrix to obtain a feature evolution probability distribution model; determining the probability value output by the feature evolution probability distribution model as a feature evolution intensity index; setting an evaluation time window using a beam search algorithm based on the feature evolution intensity index; calculating the stability score of the feature sequence within each evaluation time window; selecting the path branch that retains the optimal evolution feature based on the stability score; using the path branch as the search starting point for the next evaluation time window; determining the optimal feature evolution path based on a weighted combination of the feature evolution intensity index and the stability score of the path branch; and generating a feature evolution analysis report.

[0007] Based on the standardized feature sequence, multiple candidate feature values ​​are generated at each time point through Monte Carlo sampling. The calculation of the transition probability matrix between adjacent time points includes: performing kernel density estimation based on the standardized feature sequence to construct a feature value distribution model for each time point; setting sampling parameters according to the feature value distribution model, performing Monte Carlo sampling, and generating a preset number of candidate feature values ​​at each time point; calculating the statistical characteristics of the candidate feature values, and dynamically adjusting the sampling parameters according to the statistical characteristics until the distribution characteristics of the candidate feature values ​​and the distribution characteristics of the original feature sequence meet a similarity threshold requirement; calculating the conditional transition probability between adjacent time points based on the candidate feature values, and constructing a transition probability matrix based on the conditional transition probability.

[0008] Kernel density estimation is performed based on standardized feature sequences to construct a feature value distribution model for each time point. Sampling parameters are set according to the feature value distribution model, and Monte Carlo sampling is performed to generate a preset number of candidate feature values ​​at each time point. This includes: performing kernel density estimation based on standardized feature sequences to construct a feature value distribution model based on the feature value distribution at each time point; analyzing the probability distribution characteristics of each time point using the feature value distribution model and setting sampling parameters based on the probability distribution characteristics; and performing Monte Carlo sampling according to the sampling parameters to generate a preset number of candidate feature values ​​at each time point.

[0009] The process of iteratively optimizing the transition probability matrix to obtain a feature evolution probability distribution model, and determining the probability value output by the feature evolution probability distribution model as a feature evolution intensity index, includes: constructing an optimization objective function based on the transition probability matrix; introducing a time decay factor into the optimization objective function; iteratively optimizing the transition probability matrix according to the temporal correlation of feature value transitions to obtain the feature evolution probability distribution model; performing convergence verification on the feature evolution probability distribution model; and when the iterative optimization result meets the convergence condition, using the probability value output by the feature evolution probability distribution model as a feature evolution intensity index.

[0010] Based on the feature evolution intensity index, a beam search algorithm is used to set an evaluation time window. Within each evaluation time window, the stability score of the feature sequence is calculated. The selection of the path branch to retain the optimal evolutionary feature based on the stability score includes: setting an initial search width based on the feature evolution intensity index; traversing the feature sequence using the beam search algorithm to determine the starting position and length of the evaluation time window; extracting the change trend of the feature sequence within each evaluation time window; calculating the variance and skewness of the feature sequence to obtain a stability score; sorting the path branches of the feature sequence according to the stability score; selecting a preset number of path branches with the highest scores as retention targets; extracting evolutionary features from the retention targets; and determining the path branch with the optimal evolutionary feature.

[0011] The process of determining the optimal feature evolution path and generating a feature evolution analysis report based on a weighted combination of the feature evolution intensity index and the stability scores of the path branches includes: constructing an evaluation function based on the feature evolution intensity index and the stability scores of the path branches, and calculating a weighted combination score by dynamically adjusting the weight coefficients; sorting all path branches based on the weighted combination score and selecting the path with the highest score as the optimal feature evolution path; performing time-series analysis on the optimal feature evolution path and extracting feature change information at time nodes; identifying evolution trend inflection points based on the feature change information, determining the feature evolution stage division, and generating a feature evolution analysis report.

[0012] A second aspect of this invention provides an intelligent feature evolution analysis system for time-series paths, comprising: a module 1, configured to acquire time-series path data, preprocess the time-series path data, and generate a standardized feature sequence; a module 2, configured to generate multiple candidate feature values ​​at each time point using Monte Carlo sampling based on the standardized feature sequence, calculate a transition probability matrix between adjacent time points of the candidate feature values, perform iterative optimization based on the transition probability matrix to obtain a feature evolution probability distribution model, and determine the probability value output by the feature evolution probability distribution model as a feature evolution intensity index; a module 3, configured to set an evaluation time window using a bundle search algorithm based on the feature evolution intensity index, calculate the stability score of the feature sequence within each evaluation time window, select the path branch that retains the optimal evolution feature based on the stability score, and use the path branch as the search starting point for the next evaluation time window; and a module 4, configured to determine the optimal feature evolution path based on a weighted combination of the feature evolution intensity index and the stability score of the path branch, and generate a feature evolution analysis report.

[0013] A third aspect of the present invention provides an electronic device, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: by generating multiple candidate feature values ​​and calculating the transition probability matrix through Monte Carlo sampling, a feature evolution probability distribution model is established, which can more accurately capture the feature change patterns in time series data and avoid the bias in judging feature evolution trends in traditional methods.

[0016] By using a beam search algorithm to set the evaluation time window and selecting the path branch of the optimal evolutionary feature based on the stability score, the influence of random fluctuations in time series feature analysis is significantly reduced, ensuring the continuity and reliability of the feature evolution path.

[0017] The optimal feature evolution path is determined by a weighted combination of feature evolution intensity index and path branch stability score, realizing a multi-dimensional and multi-level feature evaluation system and improving the scientificity and rationality of feature selection.

[0018] This method significantly reduces computational complexity through iterative optimization and beam search algorithms, enabling efficient processing of large-scale time-series path data and meeting real-time analysis requirements. The generated feature evolution analysis report, based on probabilistic models and stability assessments, provides quantitative indicators and visualizations of the feature evolution process, enhancing the interpretability of the analysis results and facilitating decision support. Attached Figure Description

[0019] Figure 1 is a flowchart illustrating the intelligent feature evolution analysis method for time-series paths according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1 is a flowchart illustrating the intelligent feature evolution analysis method for time-series paths according to an embodiment of the present invention. As shown in Figure 1, the method includes: acquiring time-series path data; preprocessing the time-series path data to generate a standardized feature sequence; generating multiple candidate feature values ​​at each time point through Monte Carlo sampling based on the standardized feature sequence; calculating the transition probability matrix of candidate feature values ​​between adjacent time points; iteratively optimizing the transition probability matrix to obtain a feature evolution probability distribution model; determining the probability value output by the feature evolution probability distribution model as a feature evolution intensity index; setting an evaluation time window based on the feature evolution intensity index using a beam search algorithm; calculating the stability score of the feature sequence within each evaluation time window; selecting the path branch that retains the optimal evolution feature based on the stability score; using the path branch as the search starting point for the next evaluation time window; determining the optimal feature evolution path based on the weighted combination of the feature evolution intensity index and the stability score of the path branch; and generating a feature evolution analysis report.

[0023] In one optional implementation, based on the standardized feature sequence, generating multiple candidate feature values ​​at each time point through Monte Carlo sampling, and calculating the transition probability matrix of candidate feature values ​​between adjacent time points includes: performing kernel density estimation based on the standardized feature sequence to construct a feature value distribution model for each time point; setting sampling parameters according to the feature value distribution model, performing Monte Carlo sampling, and generating a preset number of candidate feature values ​​at each time point; calculating the statistical characteristics of the candidate feature values, and dynamically adjusting the sampling parameters according to the statistical characteristics until the distribution characteristics of the candidate feature values ​​and the distribution characteristics of the original feature sequence meet a similarity threshold requirement; calculating the conditional transition probability between adjacent time points based on the candidate feature values, and constructing a transition probability matrix based on the conditional transition probability.

[0024] Standardizing the feature sequences lays the foundation for subsequent Monte Carlo sampling. Standardization can be achieved using methods such as Z-score standardization or Min-Max standardization, which transforms the original feature sequences into standardized feature sequences with a mean of 0 and a standard deviation of 1. This effectively eliminates the influence of dimensions and makes features at different scales comparable.

[0025] Based on the standardized feature sequences, kernel density estimation is performed to construct a model of the eigenvalue distribution at each time point. Kernel density estimation is a non-parametric density estimation method that does not presuppose the data distribution form but infers the probability density function from the sample data. In practical applications, a Gaussian kernel function can be chosen. For a standardized feature sequence X_t at time point t, its kernel density estimation function can be expressed as the probability density of the eigenvalue x at that time point.

[0026] After kernel density estimation is completed, initial sampling parameters are set, including the number of samples N and the bandwidth parameter h. The number of samples is usually set to a value between 50 and 500, depending on the computational resources and accuracy requirements; the initial value of the bandwidth parameter can be set between 0.2 and 1.0, and this parameter affects the smoothness of the sampling distribution.

[0027] During Monte Carlo sampling, for each time point t, N candidate feature values ​​are generated based on the previously constructed probability density function. An acceptance-rejection sampling strategy is employed during the sampling process to ensure that the generated candidate feature values ​​conform to the original distribution characteristics.

[0028] After generating candidate feature values, their statistical properties are calculated, including mean, variance, skewness, kurtosis, and other statistics. These statistics are then compared with the corresponding statistics of the original feature sequence to calculate a similarity index. Similarity calculation can use methods such as Euclidean distance or KL divergence. If the similarity index exceeds a preset threshold (e.g., 0.9), the distribution of the candidate feature values ​​is considered sufficiently similar to the original distribution; otherwise, the sampling parameters need to be adjusted.

[0029] During parameter adjustment, if the variance of the candidate feature value distribution is less than that of the original distribution, the bandwidth parameter h is decreased; if the variance is too large, the bandwidth parameter is increased. Simultaneously, the kernel function type or sampling strategy can be adjusted based on differences in skewness and kurtosis. After parameter adjustment, Monte Carlo sampling is re-executed, and statistical properties and similarity are recalculated until the similarity threshold requirement is met or the maximum number of iterations is reached.

[0030] Once the candidate feature value distribution meets the requirements, the conditional transition probability between adjacent time points is calculated. For each candidate feature value x_i at time point t and each candidate feature value y_j at time point t+1, the conditional probability P(y_j|x_i) is calculated, representing the probability that the feature value transitions from x_i to y_j. During the calculation, the Euclidean distance and temporal correlation between feature values ​​can be considered, and a weighting function is constructed for probability allocation.

[0031] Organize all conditional probabilities into a transition probability matrix P, where the matrix element P[i, j] represents the probability of transitioning from the i-th candidate eigenvalue at time point t to the j-th candidate eigenvalue at time point t+1. The transition probability matrix is ​​an N×N matrix (N is the number of candidate eigenvalues ​​at each time point), with the sum of elements in each row being 1, satisfying the probability distribution requirements.

[0032] In practical applications, such as financial time-series data analysis, the above methods can be used to construct a transition probability matrix of stock price characteristic sequences to capture price fluctuation patterns. For example, after standardizing the daily return sequence of a stock, 100 candidate return values ​​are generated for each trading day. By calculating the transition probability between adjacent trading days, a probabilistic model reflecting market state transitions can be constructed, providing a basis for investment decisions.

[0033] Another application scenario is in industrial equipment condition monitoring, where modeling characteristic sequences such as equipment temperature and vibration is performed. By constructing an eigenvalue transition probability matrix, abnormal equipment condition transition patterns can be identified, providing early warnings of potential faults. For example, when certain feature transition probabilities are observed to deviate significantly from the normal pattern, it indicates that the equipment is about to fail, thus triggering preventative maintenance measures.

[0034] Once the transition probability matrix is ​​constructed, it can be further used for tasks such as time series feature prediction, anomaly detection, or pattern recognition, providing support for data-driven decision-making.

[0035] In one optional implementation, kernel density estimation is performed based on the standardized feature sequence to construct a feature value distribution model for each time point; sampling parameters are set according to the feature value distribution model, and Monte Carlo sampling is performed to generate a preset number of candidate feature values ​​at each time point, including: performing kernel density estimation based on the standardized feature sequence to construct a feature value distribution model according to the feature value distribution at each time point; analyzing the probability distribution characteristics of each time point using the feature value distribution model, and setting sampling parameters based on the probability distribution characteristics; performing Monte Carlo sampling according to the sampling parameters to generate a preset number of candidate feature values ​​at each time point.

[0036] Obtain the raw time series data and preprocess it, including noise removal and missing value imputation. The preprocessed data yields a complete feature sequence, denoted as X(t), where t represents a time point ranging from 1 to T.

[0037] Standardization is performed on the preprocessed feature sequences to ensure the data conforms to a distribution with a mean of 0 and a standard deviation of 1, facilitating subsequent analysis. The standardization process employs the z-score standardization method, which calculates the standardized value Z(t) for each feature value X(t) at time point t: the mean μ and standard deviation σ of the original feature sequence are calculated, and the feature values ​​at each time point t are transformed to obtain the standardized feature sequence Z(t). Standardization allows features with different dimensions to be compared and analyzed on the same scale, avoiding problems caused by inconsistencies in dimensions.

[0038] Kernel density estimation is performed based on standardized feature sequences, constructing a feature value distribution model for each time point. Kernel density estimation is a nonparametric estimation method that can learn the probability distribution from sample data without pre-assuming the distribution type. Specifically, at each time point t, historical data from the same period are collected to form a sample set {Z1(t), Z2(t), ..., Z...}. n (t)}, where n represents the number of historical samples.

[0039] A Gaussian kernel function is used for kernel density estimation. For time point t, the probability density function f(z) of its eigenvalue distribution can be expressed as the weighted sum of the kernel functions of historical samples at that point. The bandwidth parameter h of the kernel function is selected optimally using cross-validation to ensure both the smoothness of the distribution curve and the preservation of the main characteristics of the data distribution. Too small a bandwidth will lead to overfitting, while too large a bandwidth will result in excessive smoothing and loss of important distribution features.

[0040] Based on the eigenvalue distribution model obtained from kernel density estimation, the probability distribution characteristics at each time point are analyzed. The main analysis includes: the peak position of the distribution, the skewness coefficient of the distribution, the kurtosis of the distribution, the central tendency of the distribution, and the degree of dispersion. For multimodal distributions, it is also necessary to identify each peak and its corresponding probability density.

[0041] Monte Carlo sampling parameters are set based on the characteristics of probability distribution, and appropriate sampling algorithms are selected according to the complexity of the distribution. For unimodal and approximately normal distributions, the accept-reject sampling method can be used directly; for multimodal or asymmetric distributions, Markov chain Monte Carlo methods, such as the Metropolis-Hastings algorithm, are used.

[0042] The sampling parameters mainly include: sampling algorithm type, proposal distribution parameters, acceptance probability threshold, Markov chain length, and warm-up period length. For the eigenvalue distribution at time point t, if it exhibits significant asymmetry or multimodal characteristics, the Markov chain length and warm-up period should be increased to ensure sufficient mixing of samples; if the distribution is close to normal, the sampling parameter configuration can be appropriately simplified.

[0043] Based on the set sampling parameters, Monte Carlo sampling is performed. For each time point t, according to its eigenvalue distribution model f(z), a preset number M candidate eigenvalues ​​{z̃1(t), z̃2(t), ..., z̃ n The sampling process needs to ensure that the generated samples accurately reflect the characteristics of the original distribution, especially capturing the tail behavior and multimodal properties of the distribution.

[0044] When using the Metropolis-Hastings algorithm, an initial state z0 is chosen. Then, at each step, a new candidate state z' is generated using the proposal distribution. The acceptance probability α is calculated, and a decision is made on whether to accept the candidate state. This process is repeated until a sufficient number of samples are generated. To eliminate the influence of the initial state, samples generated during the warm-up period are discarded.

[0045] The generated candidate feature values ​​are transformed back to the original feature space through inverse normalization, that is, the candidate feature values ​​z̃ for time point t. i (t), calculate its corresponding original feature space value X̃ i (t). These candidate eigenvalues ​​constitute a probabilistic prediction of future eigenvalues.

[0046] The quality of the generated candidate feature values ​​is evaluated by calculating the statistical properties of the candidate feature value set, such as mean, variance, and quantiles, and comparing them with the corresponding statistics of the original distribution to ensure that the generated candidate feature value set can accurately reflect the characteristics of the original distribution.

[0047] In practical applications, such as industrial equipment condition monitoring, the above methods can be used to predict the range of equipment parameters and thus assess the risk of equipment failure. By performing kernel density estimation and Monte Carlo sampling on historical sensor data, various scenarios of the equipment's future state can be generated, providing decision support for preventative maintenance.

[0048] The advantage of this method is that it does not require assumptions about the data distribution type, can adaptively learn complex distributions, and generate candidate values ​​that conform to historical distribution characteristics through Monte Carlo sampling, providing reliable probabilistic model support for subsequent risk assessment and decision-making.

[0049] In one optional implementation, the feature evolution probability distribution model is obtained by iterative optimization based on the transition probability matrix, and the probability value output by the feature evolution probability distribution model is determined as the feature evolution intensity index. This includes: constructing an optimization objective function based on the transition probability matrix, introducing a time decay factor into the optimization objective function, iteratively optimizing the transition probability matrix according to the temporal correlation of feature value transitions to obtain the feature evolution probability distribution model; performing convergence verification on the feature evolution probability distribution model, and when the iterative optimization result meets the convergence condition, using the probability value output by the feature evolution probability distribution model as the feature evolution intensity index.

[0050] The transition probability matrix reflects the evolution of time-series features between adjacent time points. The matrix has an N×N dimension, where N is the number of discretized intervals for the eigenvalues. Each element Pij in the matrix represents the conditional probability of an eigenvalue transitioning from interval i to interval j, calculated as the co-occurrence frequency of eigenvalues ​​in intervals i and j divided by the total frequency of eigenvalues ​​in interval i. To improve the reliability of the probability estimation, Laplace smoothing is employed, with a smoothing factor α (value 0.01) added to each probability.

[0051] The objective function J(P) is constructed based on the transition probability matrix, consisting of three terms: a transition probability term Jp, a time correlation term Jt, and a regularization term Jr. The transition probability term Jp measures the log-likelihood of eigenvalue transitions, the time correlation term Jt introduces a time decay factor to attenuate long-term effects, and the regularization term Jr prevents overfitting. The objective function is expressed as: the sum of the logarithms of the transition probabilities at all time points in the feature sequence, multiplied by the corresponding time decay factor, and then the L2 regularization term is added.

[0052] The time decay factor adopts the exponential decay function exp(-λd), where d is the time interval and λ is the decay coefficient. The optimal value of the decay coefficient λ is determined through cross-validation, with a validation interval of [0.05, 0.5] and a step size of 0.05. A larger λ value makes the model focus more on short-term evolutionary features, while a smaller λ value retains more long-term evolutionary information. Practice shows that a value of λ in the range of 0.1-0.2 is more suitable.

[0053] The iterative optimization process employs the stochastic gradient descent algorithm. In each iteration, a mini-batch (size 32) is randomly selected to calculate the gradient, and the Adam optimizer is used to update the parameters. The hyperparameters of the Adam optimizer are set as follows: initial learning rate 0.01, β1=0.9, β2=0.999, ε=1e-8. To prevent gradient vanishing, the gradients from each iteration are normalized. After each parameter update, the transition probability matrix is ​​row-normalized to ensure that the sum of the probabilities is 1.

[0054] Convergence verification employs a multi-indicator joint judgment strategy, calculating the maximum relative error (max_err), the average relative error (mean_err), and the relative change in the objective function value (obj_diff) of the transition probability matrix elements between adjacent iterations. Three thresholds are set: max_err_thresh = 0.001, mean_err_thresh = 0.0005, and obj_diff_thresh = 0.0001. If all three indicators are less than the corresponding thresholds for 10 consecutive iterations, the iteration is considered converged. If convergence is not achieved after reaching the maximum number of iterations (default 1000), the current optimal result is returned.

[0055] When converting the output probability values ​​of the feature evolution probability distribution model into a feature evolution intensity index, multiple statistical features are considered. For each time point t, the transition probability distribution at that point is extracted, and four statistics are calculated: the probability mean (mean_p), the probability variance (var_p), the probability entropy (entropy_p), and the maximum probability (max_p). These statistics are fused using a weighted combination method, with the weights obtained through grid search optimization. The default weight configuration is: mean_p (0.4), var_p (0.2), entropy_p (0.2), and max_p (0.2).

[0056] Taking a stock price trend sequence as an example in practical applications, the intraday highest price data of a stock over 300 trading days is selected. The price data is normalized to the [0,1] interval, and the feature values ​​are discretized into 20 equal-width intervals. A 20×20 transition probability matrix is ​​constructed, and the time decay coefficient λ=0.15 is set. Iterative optimization is performed using the above parameter configuration. After 326 iterations, the convergence condition is reached. In the final feature evolution probability distribution model, the average transition probability between adjacent time points is 0.283, and the statistics are: mean_p=0.283, var_p=0.042, entropy_p=2.17, max_p=0.412. The calculated feature evolution strength index is 0.375, indicating that the feature sequence exhibits moderately weak evolutionary characteristics.

[0057] To ensure model stability, an early stopping strategy and a model checkpoint mechanism are introduced. Model performance is evaluated on the validation set every 50 iterations, and iteration is terminated early if performance fails to improve after three consecutive iterations. Simultaneously, optimal model parameters are saved during iteration, and if numerical instability occurs, the model can revert to the optimal checkpoint. For online applications, a sliding window approach is used to periodically update the model, with a default window size of 90 time points and a step size of 30 time points.

[0058] In one optional implementation, based on the feature evolution intensity index, an evaluation time window is set using a beam search algorithm. Within each evaluation time window, a stability score of the feature sequence is calculated. Selecting the path branch that retains the optimal evolutionary feature based on the stability score includes: setting an initial search width based on the feature evolution intensity index; traversing the feature sequence using the beam search algorithm to determine the starting position and length of the evaluation time window; extracting the change trend of the feature sequence within each evaluation time window; calculating the variance and skewness of the feature sequence to obtain a stability score; sorting the path branches of the feature sequence according to the stability score; selecting a preset number of path branches with the highest scores as retention targets; extracting evolutionary features from the retention targets; and determining the path branch with the optimal evolutionary feature.

[0059] Acquire the feature sequence data to be analyzed. This data contains feature values ​​at multiple time points. The feature sequence can be user behavior data, device operating parameters, market indicators, or other quantitative indicators that change over time. Preprocess the acquired feature sequence data, including data cleaning, normalization, and missing value imputation, to ensure the accuracy of subsequent analysis.

[0060] The Feature Evolution Intensity Index (FEI) is used to quantify the degree of change of a feature in a time series. The calculation method is as follows: perform time difference operations on the feature sequence to obtain the change between adjacent time points; calculate the average of the absolute values ​​of these changes as the FEI value. A larger FEI value indicates more drastic feature change; a smaller value indicates more gradual feature change.

[0061] Based on the calculated feature evolution intensity index, an initial search width is set. The initial search width is positively correlated with the feature evolution intensity; that is, the more drastic the feature change, the more candidate paths need to be retained to ensure that potential optimal paths are not missed. Specifically, the initial search width can be set as the logarithm of the evolution intensity multiplied by a preset coefficient and rounded up. For example, when the feature evolution intensity index is 4.5, if the preset coefficient is 2, the initial search width is 9.

[0062] A beam search algorithm is employed to traverse the feature sequence. Beam search is a heuristic search algorithm that retains a fixed number of optimal candidates at each step to balance search efficiency and result quality. In this method, beam search is used to determine the starting position and length of the evaluation time window. The specific steps are as follows: Starting from the beginning position of the feature sequence, consider time windows of different lengths, ranging from a minimum to a maximum length, such as 3 to 15 time units. For each combination of starting position and window length, a candidate evaluation time window is formed.

[0063] For each candidate evaluation time window, the trend of feature sequence changes within that window is extracted. The method for extracting the trend includes: calculating the first-order difference sequence of feature values ​​within the window; and using linear regression to fit the original feature values ​​to obtain the slope as an overall trend indicator.

[0064] The variance of the feature sequence within the window is calculated. Variance reflects the magnitude of fluctuations in the feature values; the smaller the variance, the more stable the feature sequence. The variance calculation formula uses standard statistical methods to calculate the average of the squared deviations from the mean for all feature values ​​within the window.

[0065] The skewness of the feature sequence within the window is calculated. Skewness measures the asymmetry of the data distribution and is used to assess the degree of balance in feature variation. The smaller the absolute value of the skewness, the more balanced the feature variation.

[0066] The stability score is calculated by combining variance and skewness. The formula for the stability score is: Stability Score = 1 / (Variance * (1 + Absolute value of skewness)). This formula ensures that windows with small variance and near-zero skewness achieve a high stability score.

[0067] Based on the calculated stability scores, all candidate path branches are sorted from highest to lowest. A predetermined number of path branches with the highest stability scores are selected as retained targets; this predetermined number is typically equal to the initial search width.

[0068] Evolutionary feature extraction is performed on the retained path branches. The evolutionary feature extraction includes the following steps: For each retained path branch, calculate the dynamic characteristics such as the growth rate, fluctuation cycle, and mutation point of its features within the window; extract the trend component, seasonal component, and random component of the feature sequence and analyze the changes in the proportion of each component; identify the patterns and regularities in the feature sequence, such as repetitive patterns and increasing / decreasing trends.

[0069] The optimal evolutionary feature path branch is determined by comparing the completeness, representativeness, and predictive power of the evolutionary features extracted from each path branch. Completeness refers to whether the features cover the main changes in the sequence; representativeness refers to whether the features accurately reflect the essential characteristics of the sequence; and predictive power refers to the accuracy of predictions based on the extracted features.

[0070] The final output includes the evaluation time window parameters and the corresponding set of evolutionary features for the optimal path branch, serving as the basis for subsequent analysis and decision-making. These evolutionary features can be used in applications such as predicting future trends, identifying abnormal behavior, or optimizing system parameters.

[0071] The above methods can effectively capture key evolutionary features in feature sequences, improve the accuracy and efficiency of data analysis, and provide reliable support for various applications that require time series analysis.

[0072] In one optional implementation, determining the optimal feature evolution path and generating a feature evolution analysis report based on a weighted combination of the feature evolution intensity index and the stability scores of the path branches includes: constructing an evaluation function based on the feature evolution intensity index and the stability scores of the path branches, and calculating a weighted combination score by dynamically adjusting the weight coefficients; sorting all path branches based on the weighted combination score, and selecting the path with the highest score as the optimal feature evolution path; performing time-series analysis on the optimal feature evolution path to extract feature change information at time nodes; identifying evolution trend inflection points based on the feature change information, determining the feature evolution stage division, and generating a feature evolution analysis report.

[0073] An evaluation function is constructed to comprehensively consider the feature evolution strength index and the path branch stability score. This evaluation function can be expressed as a linear combination of the feature evolution strength index and the path branch stability score. The feature evolution strength index reflects the degree and importance of feature changes during evolution, while the path branch stability score reflects the ability of a path to maintain consistency under different conditions.

[0074] The weight coefficients in the evaluation function are determined through a dynamic adjustment mechanism that automatically optimizes the weight allocation based on historical evolution data and the current application scenario. Specifically, a genetic algorithm is used to iteratively optimize the weight parameters, with initial weights set to α=0.6 (feature evolution strength) and β=0.4 (path stability). During the iteration process, the weights are adjusted based on the prediction accuracy on the validation set until convergence or the preset number of iterations is reached.

[0075] After the weights are determined, the weighted combination score of each path branch is calculated. For example, for a risk prediction model in the financial field, if the characteristic evolution intensity index of a specific path is 0.85 and the stability score is 0.72, then its weighted combination score is 0.85×0.6+0.72×0.4=0.798.

[0076] All path branches are sorted based on weighted combined scores using a quicksort algorithm to ensure efficiency on large-scale path sets. After sorting, the path with the highest score is selected as the optimal feature evolution path. If multiple paths with similar scores exist (difference less than a preset threshold of 0.05), the complexity and interpretability of the paths are further examined, and the path with lower complexity and higher interpretability is selected.

[0077] The next key step is to perform time-series analysis on the optimal feature evolution path. Time-series analysis uses a sliding window technique to extract feature change information at time nodes. The window size is dynamically adjusted according to the data characteristics, typically 5%-15% of the total time span. Within each time window, the Euclidean distance or cosine similarity of the feature vectors is calculated to quantify the degree of feature change between adjacent time windows.

[0078] Based on the extracted feature change information, inflection points in the evolutionary trend are identified. Inflection point identification employs a curvature-based calculation method combined with Z-score anomaly detection technology. When the feature change rate exceeds two standard deviations from the historical average, the point is marked as a potential inflection point. Simultaneously, the DTW (Dynamic Time Warping) algorithm is used to compare the feature change sequence with typical patterns to verify the effectiveness of the inflection points.

[0079] After identifying the inflection point, the feature evolution stages are divided. The division process considers temporal continuity and feature similarity, and similar feature change stages are merged using a hierarchical clustering algorithm. For example, in e-commerce user behavior analysis, it is divided into four stages: "initial exploration period," "rapid growth period," "stable usage period," and "decline and transformation period."

[0080] Generate a feature evolution analysis report, which includes a visualization of the optimal feature evolution path, feature descriptions of each evolution stage, inflection point event analysis, and future trend predictions. The report adopts a multi-level structure, unfolding layer by layer from overall trends to specific feature changes, and provides different levels of detail for different audiences (such as technical analysts and decision-makers).

[0081] In practical applications, taking social media user behavior analysis as an example, the above methods can accurately identify the key inflection point in the shift from "content browsing-driven" to "content creation-driven" users, and predict the evolution trend of user activity, providing data support for platform function optimization and user retention strategies.

[0082] The entire process forms a closed-loop feedback mechanism. As new data accumulates, the weight coefficients of the evaluation function are updated periodically to ensure the accuracy and timeliness of evolutionary path analysis. In this way, the optimal selection of feature evolution paths and the automatic generation of analysis reports are achieved, providing reliable support for data-driven decision-making.

[0083] A second aspect of this invention provides an intelligent feature evolution analysis system for time-series paths, comprising: a module 1, configured to acquire time-series path data, preprocess the time-series path data, and generate a standardized feature sequence; a module 2, configured to generate multiple candidate feature values ​​at each time point using Monte Carlo sampling based on the standardized feature sequence, calculate a transition probability matrix between adjacent time points of the candidate feature values, perform iterative optimization based on the transition probability matrix to obtain a feature evolution probability distribution model, and determine the probability value output by the feature evolution probability distribution model as a feature evolution intensity index; a module 3, configured to set an evaluation time window using a bundle search algorithm based on the feature evolution intensity index, calculate the stability score of the feature sequence within each evaluation time window, select the path branch that retains the optimal evolution feature based on the stability score, and use the path branch as the search starting point for the next evaluation time window; and a module 4, configured to determine the optimal feature evolution path based on a weighted combination of the feature evolution intensity index and the stability score of the path branch, and generate a feature evolution analysis report.

[0084] A third aspect of the present invention provides an electronic device, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the aforementioned method.

[0085] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0086] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent feature evolution analysis of temporal paths, characterized in that, include: Acquire time-series path data, preprocess the time-series path data, and generate standardized feature sequences; Based on the standardized feature sequence, multiple candidate feature values ​​are generated at each time point through Monte Carlo sampling. The transition probability matrix of candidate feature values ​​between adjacent time points is calculated. Based on the transition probability matrix, an iterative optimization is performed to obtain a feature evolution probability distribution model. The probability value output by the feature evolution probability distribution model is determined as the feature evolution intensity index. Based on the feature evolution intensity index, a beam search algorithm is used to set an evaluation time window. Within each evaluation time window, the stability score of the feature sequence is calculated. Based on the stability score, the path branch that retains the optimal evolution feature is selected, and the path branch is used as the search starting point for the next evaluation time window. Based on the weighted combination of the feature evolution intensity index and the stability score of the path branch, the optimal feature evolution path is determined, and a feature evolution analysis report is generated.

2. The method according to claim 1, characterized in that, Based on the standardized feature sequence, multiple candidate feature values ​​are generated at each time point through Monte Carlo sampling. The calculation of the transition probability matrix between adjacent time points includes: performing kernel density estimation based on the standardized feature sequence to construct a feature value distribution model for each time point; setting sampling parameters according to the feature value distribution model, performing Monte Carlo sampling, and generating a preset number of candidate feature values ​​at each time point; calculating the statistical characteristics of the candidate feature values, and dynamically adjusting the sampling parameters according to the statistical characteristics until the distribution characteristics of the candidate feature values ​​and the distribution characteristics of the original feature sequence meet a similarity threshold requirement; calculating the conditional transition probability between adjacent time points based on the candidate feature values, and constructing a transition probability matrix based on the conditional transition probability.

3. The method according to claim 2, characterized in that, Kernel density estimation is performed based on standardized feature sequences to construct a feature value distribution model for each time point. Sampling parameters are set according to the feature value distribution model, and Monte Carlo sampling is performed to generate a preset number of candidate feature values ​​at each time point. This includes: performing kernel density estimation based on standardized feature sequences to construct a feature value distribution model based on the feature value distribution at each time point; analyzing the probability distribution characteristics of each time point using the feature value distribution model and setting sampling parameters based on the probability distribution characteristics; and performing Monte Carlo sampling according to the sampling parameters to generate a preset number of candidate feature values ​​at each time point.

4. The method according to claim 1, characterized in that, The process of iteratively optimizing the transition probability matrix to obtain a feature evolution probability distribution model, and determining the probability value output by the feature evolution probability distribution model as a feature evolution intensity index, includes: constructing an optimization objective function based on the transition probability matrix; introducing a time decay factor into the optimization objective function; iteratively optimizing the transition probability matrix according to the temporal correlation of feature value transitions to obtain the feature evolution probability distribution model; performing convergence verification on the feature evolution probability distribution model; and when the iterative optimization result meets the convergence condition, using the probability value output by the feature evolution probability distribution model as a feature evolution intensity index.

5. The method according to claim 1, characterized in that, Based on the feature evolution intensity index, a beam search algorithm is used to set an evaluation time window. Within each evaluation time window, the stability score of the feature sequence is calculated. The selection of the path branch to retain the optimal evolutionary feature based on the stability score includes: setting an initial search width based on the feature evolution intensity index; traversing the feature sequence using the beam search algorithm to determine the starting position and length of the evaluation time window; extracting the change trend of the feature sequence within each evaluation time window; calculating the variance and skewness of the feature sequence to obtain a stability score; sorting the path branches of the feature sequence according to the stability score; selecting a preset number of path branches with the highest scores as retention targets; extracting evolutionary features from the retention targets; and determining the path branch with the optimal evolutionary feature.

6. The method according to claim 1, characterized in that, The process of determining the optimal feature evolution path and generating a feature evolution analysis report based on a weighted combination of the feature evolution intensity index and the stability scores of the path branches includes: constructing an evaluation function based on the feature evolution intensity index and the stability scores of the path branches, and calculating a weighted combination score by dynamically adjusting the weight coefficients; sorting all path branches based on the weighted combination score and selecting the path with the highest score as the optimal feature evolution path; performing time-series analysis on the optimal feature evolution path and extracting feature change information at time nodes; identifying evolution trend inflection points based on the feature change information, determining the feature evolution stage division, and generating a feature evolution analysis report.

7. A time-series path intelligent feature evolution analysis system, used to implement the method of any one of claims 1-6, characterized in that, include: Module 1 is used to acquire time-series path data, preprocess the time-series path data, and generate standardized feature sequences. Module 2 is used to generate multiple candidate feature values ​​at each time point based on the standardized feature sequence through Monte Carlo sampling, calculate the transition probability matrix of candidate feature values ​​between adjacent time points, and perform iterative optimization based on the transition probability matrix to obtain a feature evolution probability distribution model. The probability value output by the feature evolution probability distribution model is determined as the feature evolution intensity index. Module 3 is used to set an evaluation time window based on the feature evolution intensity index using a bundle search algorithm, calculate the stability score of the feature sequence within each evaluation time window, select the path branch that retains the optimal evolution feature based on the stability score, and use the path branch as the search starting point for the next evaluation time window. Module 4 is used to determine the optimal feature evolution path based on the weighted combination of the feature evolution intensity index and the stability score of the path branch, and generate a feature evolution analysis report.

8. An electronic device, characterized in that, include: processor; A memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images